Inline Float64x2 methods that take 1 argument

runtime/vm/intermediate_language_arm.cc

Index: runtime/vm/intermediate_language_arm.cc
diff --git a/runtime/vm/intermediate_language_arm.cc b/runtime/vm/intermediate_language_arm.cc
index dcd110e3926bb77706105329322c82fef1f394d6..f1d5b91b5d36447e6ee93f1005aece56ab06c386 100644
--- a/runtime/vm/intermediate_language_arm.cc
+++ b/runtime/vm/intermediate_language_arm.cc
@@ -4093,6 +4093,60 @@ void Float64x2ZeroArgInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
 }
 
 
+LocationSummary* Float64x2OneArgInstr::MakeLocationSummary(bool opt) const {
+  const intptr_t kNumInputs = 2;
+  const intptr_t kNumTemps = 0;
+  LocationSummary* summary =
+      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
+  summary->set_in(0, Location::RequiresFpuRegister());
+  summary->set_in(1, Location::RequiresFpuRegister());
+  summary->set_out(Location::SameAsFirstInput());
+  return summary;
+}
+
+
+void Float64x2OneArgInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  QRegister left = locs()->in(0).fpu_reg();
+  DRegister left0 = EvenDRegisterOf(left);
+  DRegister left1 = OddDRegisterOf(left);
+  QRegister right = locs()->in(1).fpu_reg();
+  DRegister right0 = EvenDRegisterOf(right);
+  DRegister right1 = OddDRegisterOf(right);
+  QRegister out = locs()->out().fpu_reg();
+  ASSERT(left == out);
+
+  switch (op_kind()) {
+    case MethodRecognizer::kFloat64x2Scale:
+      __ vmuld(left0, left0, right0);
+      __ vmuld(left1, left1, right0);
+      break;
+    case MethodRecognizer::kFloat64x2WithX:
+      __ vmovd(left0, right0);
+      break;
+    case MethodRecognizer::kFloat64x2WithY:
+      __ vmovd(left1, right0);
+      break;
+    case MethodRecognizer::kFloat64x2Min:
+      __ vcmpd(left0, right0);
+      __ vmstat();
+      __ vmovd(left0, right0, GT);

[zra, 2014/03/10 22:14:08]

I double checked the performance guide, and it looks like it recommends
preferring branches instead of conditional FPU instructions. If it's not too
much trouble, why don't we go with that first? Something like:

Label le0;
__ vcmpd(left0, right0);
__ vmstat();
__ b(&le0, LE);
__ vmovd(left0, right0);
__ Bind(&le0)

[Cutch, 2014/03/10 22:26:45]

Done.

+      __ vcmpd(left1, right1);
+      __ vmstat();
+      __ vmovd(left1, right1, GT);
+      break;
+    case MethodRecognizer::kFloat64x2Max:
+      __ vcmpd(left0, right0);
+      __ vmstat();
+      __ vmovd(left0, right0, LT);

[zra, 2014/03/10 22:14:08]

Same here.

[Cutch, 2014/03/10 22:26:45]

Done.

+      __ vcmpd(left1, right1);
+      __ vmstat();
+      __ vmovd(left1, right1, LT);
+      break;
+    default: UNREACHABLE();
+  }
+}
+
+
 LocationSummary* Int32x4BoolConstructorInstr::MakeLocationSummary(
     bool opt) const {
   const intptr_t kNumInputs = 4;